How do I do it?
We use a novel context-aware neural-network-based pairwise ranker to improve Search Suggestion Ranking for providing the users with the top âkâ ranked suggestions out of a corpus of ânâ possible suggestions based on the prefix of the query that has been entered so far.
There are 2 main components in this- Retrieval component & Reranking component. Retrieval component: This component fetches the candidate suggestions for a query prefix. Reranking component (Neural-network based pairwise reranker): This component reranks the candidate suggestions given a query prefix in accordance with the relevance.
Denoising autoencoder trained as a Siamese network with a dedicated decoder for query and suggestion reconstruction enabled the model to learn data distribution more intuitively. Model is evaluated on MRR, CTR, query length as primary metrics. The model showed significant improvement over these evaluation metrics. The model gives embeddings at various levels which can be used for various downstream tasks.